agent-setup

My AI coding-agent toolkit. One script reproduces the full setup on any computer, for Claude Code, Codex, OpenCode, and Antigravity.

What it installs

Layer	Tool	What it does
tokless	caveman	Terse output - ~65% fewer tokens
	ponytail	Lazy-senior-dev minimalism, no over-engineering
	rtk	Compresses command output before it hits the LLM
	codegraph	Pre-indexed code knowledge graph (MCP)
	context-mode	Sandboxed tool output + session memory (MCP)
ECC	minimal + core profiles	Skills, rules, agents, commands (/plan, /code-review, /learn, ...)
impeccable	design skill	23 design commands + detector rules
karpathy-guidelines	vendored	Always-on coding discipline: think first, simplify, surgical edits, verifiable goals
ponytail	vendored	Always-on lazy-senior-dev minimalism, plus 5 one-shot report commands
Codex	.codex/ + codex/	Portable config, multi-agent roles, MCP baseline, and vendored skills
OpenCode	opencode.json + .opencode/	Project MCP baseline, agents, rtk plugin, and Codex-skill sync
Personal	claude/skills/	12 design skills (brandkit, gpt-taste, minimalist-ui, ...) + always-on skills
Personal	claude/CLAUDE.md	Global instructions for all projects

Install (new computer)

Prereqs: Node.js + npm. Agents you use (Claude Code, Codex, OpenCode) installed first.

Windows (PowerShell):

git clone https://github.com/iamtyroon/agent-setup
cd agent-setup
.\install.ps1

macOS / Linux:

git clone https://github.com/iamtyroon/agent-setup
cd agent-setup
./install.sh

Then restart your agent sessions.

Update on an existing computer

tokless update
ecc auto-update
git pull && ./install.sh   # or .\install.ps1

Repo layout

install.ps1        Windows installer
install.sh         macOS/Linux installer
AGENTS.md          Repo instructions loaded by Codex
.codex/config.toml Project-local Codex baseline and portable MCP servers
.codex/agents/     Codex multi-agent role layers
codex/skills/      Vendored Codex skills copied to the global Codex skill path
codex/references/  Shared references for the vendored OpenAI game-studio slice
codex/AGENTS.md    Optional global Codex guidance
opencode.json      Project-local OpenCode MCP and skill-permission baseline
.opencode/agents/  OpenCode subagents mirroring the Codex baseline roles
.opencode/plugins/ OpenCode-native rtk command-rewrite plugin
scripts/           Cross-platform setup helpers
claude/CLAUDE.md   Global Claude instructions
claude/skills/     Personal skills (copied to ~/.claude/skills)

Codex and OpenCode config, agent roles, skill details, and source links live in codex/README.md and opencode/README.md.

ponytail and karpathy-guidelines are vendored under both codex/skills/ and claude/skills/, so one install reaches Codex, OpenCode, and Claude Code. The installer runs tokless wiring for Codex.

In Codex, caveman and i-have-adhd are automatic defaults. Their short rules also live in AGENTS.md, so behavior does not depend on a skill command. karpathy-guidelines and ponytail are always-on: load both at the start of any coding task. karpathy-guidelines decides scope, ponytail decides size; on conflict take the smaller diff. Relevant skills autoload from their descriptions. grill-me, user-context, share, and yeet remain explicit-only because they are intrusive or change state. ponytail-audit, ponytail-debt, ponytail-gain, ponytail-help, and ponytail-review are one-shot reports, not edits, so they stay explicit-only too. unslop and impeccable activate when the task matches their descriptions or when you invoke them directly.

Safe to re-run: every step is idempotent.

The Codex config sync is add-only. It adds missing MCP, feature, and agent sections to $CODEX_HOME/config.toml; it preserves existing sections and credentials. Existing global AGENTS.md and agent role files are preserved.

OpenCode support

Run OpenCode from this repository root. opencode.json provides the portable GitHub, Context7, Exa, Memory, Playwright, Sequential Thinking, and Godot MCP baseline. tokless continues to wire machine-local codegraph and context-mode entries. .opencode/plugins/rtk.ts rewrites supported shell commands through the installed rtk binary.

The installer runs scripts/sync-opencode.mjs. It flattens every valid codex/skills/**/SKILL.md bundle, plus valid skills found in the local Codex skill and plugin caches, into OpenCode's user-wide skill directory. It also converts Codex TOML roles into OpenCode Markdown agents. Existing OpenCode agents and configuration are preserved; managed skill files are refreshed.

OpenCode session history, project registry, credentials, plugin cache, and MCP cache stay in user-local application data. They are intentionally not committed to this repository.

Codex remote plugins do not have a universal OpenCode plugin format. Their portable skill payloads and MCP baseline are synced here; OpenCode-native plugins remain user-installed in OpenCode's global config.

Notes / known quirks

• tokless binary locked: if tokless is already installed, its binary may be locked by a running agent process, so the installer skips the download. Upgrade with tokless update after closing agent sessions.
• Step order matters: ECC and impeccable installs can overwrite agent MCP config, so tokless wiring runs last. If tokless doctor ever reports missing tools after another install, re-run tokless --agents claude,opencode,codex,antigravity.
• ecc doctor warnings: "drifted managed files" warnings are expected — tokless and impeccable touch ECC-managed configs. Not an error; the installer treats them as non-fatal.
• Project-local artifacts: impeccable installs into the current directory too; .gitignore keeps those (.claude/, .codegraph/, etc.) out of the repo.